Abstract
Information Extraction (IE) currently exists in two distinct paradigms: "Closed" systems like DBpedia that utilize strict, logical schemas, and "Open" systems like NELL that extract massive but noisy datasets from the web. This paper proposes a framework to bridge these worlds by linking ambiguous Open IE strings to unique identifiers in Closed IE systems. By leveraging Wikipedia anchor text as a probabilistic bridge, the authors demonstrate a baseline for entity linking that achieves high precision and recall, providing a foundation for turning web-scale data into structured, machine-readable knowledge.
Introduction
The field of Information Extraction is divided between the precision of structured ontologies and the scale of web-based extraction. Closed IE systems (e.g., DBpedia, YAGO) are highly accurate but limited to encyclopedic domains. Conversely, Open IE systems (e.g., NELL, ReVerb) offer massive coverage but produce ambiguous "noisy" data. For instance, an Open IE system might extract a triple like (royals, plays, mlb), but without entity linking, it remains unclear if "royals" refers to the Kansas City Royals baseball team, a family of monarchs, or a financial institution. This paper addresses the challenge of Entity Linking: mapping ambiguous strings from Open IE to unique URIs in a Closed IE system to enable machine reasoning and improved search utility.
Methodology
The authors propose a most-frequent-entity baseline that utilizes Wikipedia as a "Rosetta Stone" for disambiguation. The core of this approach is the Anchor Text Bridge. Wikipedia articles contain internal links where the clickable text (the "anchor") leads to a specific page (the "entity").
By analyzing Wikipedia dumps, the system calculates the conditional probability of an entity given an anchor string. For example, as shown in the table below, the anchor "jaguar" most frequently links to "Jaguar Cars," followed by "Jaguar Racing" and the animal "Jaguar."
For any given term in an Open IE triple (subject or object), the system generates a ranked list of the top $k$ most likely DBpedia matches based on these global link frequencies. Disambiguation is then performed by selecting the candidate that best fits the context of the extracted fact.
Experimental Setup
The study utilizes several key resources:
- NELL (Never-Ending Language Learning): The source of the "Open" triples.
- DBpedia: The target "Closed" ontology used for structured URIs.
- WikiPrep: A preprocessing tool used to extract link counts from Wikipedia dumps.
A major contribution of this work is the creation of a manually annotated "Gold Standard" benchmark. The authors sampled 1,200 triples from 12 different NELL predicates (e.g., bookwriter, lakeinstate) and had human annotators map them to DBpedia to create a ground truth for evaluation.
Results and Analysis
The baseline approach performed remarkably well for a frequency-based model, achieving a micro-averaged precision of 82.78% and a recall of 81.31% for top-1 matches. However, performance varied significantly across different predicates.
As illustrated in the figure above, predicates like lakeinstate achieved the highest precision (90%) due to the relative specificity of lake names. In contrast, personleadsorganization proved much harder (precision 71%, recall 58%) due to the high ambiguity of human names. The authors also identified an "unmatchable" problem: approximately 15% of entities in certain categories (like agentcollaborateswithagent) could not be linked because they were too vague—often providing only a first name or surname—or were non-existent in the structured database.
Conclusion
This work demonstrates that a simple probabilistic baseline using Wikipedia anchor text can effectively link Open IE data to structured ontologies. While successful, the current model relies on global frequency and lacks semantic context; it does not yet "reason" that a bankbankincountry predicate should prioritize financial institutions over sports teams. Future research will focus on incorporating string similarity measures (like Levenshtein distance) and developing more complex data-driven models to handle the inherent noise of the web more robustly.
Summary updated